Merge tag 'kvm-riscv-6.8-2' of https://github.com/kvm-riscv/linux into HEAD
[linux-block.git] / kernel / dma / debug.c
CommitLineData
45051539 1// SPDX-License-Identifier: GPL-2.0-only
f2f45e5f
JR
2/*
3 * Copyright (C) 2008 Advanced Micro Devices, Inc.
4 *
5 * Author: Joerg Roedel <joerg.roedel@amd.com>
f2f45e5f
JR
6 */
7
f737b095
RM
8#define pr_fmt(fmt) "DMA-API: " fmt
9
68db0cf1 10#include <linux/sched/task_stack.h>
972aa45c 11#include <linux/scatterlist.h>
a1fd09e8 12#include <linux/dma-map-ops.h>
29930025 13#include <linux/sched/task.h>
6c132d1b 14#include <linux/stacktrace.h>
30dfa90c 15#include <linux/spinlock.h>
b4a0f533 16#include <linux/vmalloc.h>
788dcfa6 17#include <linux/debugfs.h>
8a6fc708 18#include <linux/uaccess.h>
23a7bfae 19#include <linux/export.h>
2d62ece1 20#include <linux/device.h>
f2f45e5f 21#include <linux/types.h>
2d62ece1 22#include <linux/sched.h>
8a6fc708 23#include <linux/ctype.h>
f2f45e5f 24#include <linux/list.h>
6bf07871 25#include <linux/slab.h>
2e34bde1 26#include <asm/sections.h>
a1fd09e8 27#include "debug.h"
2e34bde1 28
5e76f564 29#define HASH_SIZE 16384ULL
30dfa90c
JR
30#define HASH_FN_SHIFT 13
31#define HASH_FN_MASK (HASH_SIZE - 1)
32
15b28bbc 33#define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
2b9d9ac0 34/* If the pool runs out, add this many new entries at once */
ad78dee0 35#define DMA_DEBUG_DYNAMIC_ENTRIES (PAGE_SIZE / sizeof(struct dma_debug_entry))
15b28bbc 36
f2f45e5f
JR
37enum {
38 dma_debug_single,
f2f45e5f
JR
39 dma_debug_sg,
40 dma_debug_coherent,
0e74b34d 41 dma_debug_resource,
f2f45e5f
JR
42};
43
6c9c6d63
SK
44enum map_err_types {
45 MAP_ERR_CHECK_NOT_APPLICABLE,
46 MAP_ERR_NOT_CHECKED,
47 MAP_ERR_CHECKED,
48};
49
6c132d1b
DW
50#define DMA_DEBUG_STACKTRACE_ENTRIES 5
51
0abdd7a8
DW
52/**
53 * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
54 * @list: node on pre-allocated free_entries list
55 * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
479623fd 56 * @dev_addr: dma address
0abdd7a8 57 * @size: length of the mapping
d3694f30 58 * @type: single, page, sg, coherent
0abdd7a8
DW
59 * @direction: enum dma_data_direction
60 * @sg_call_ents: 'nents' from dma_map_sg
61 * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
d3694f30
ED
62 * @pfn: page frame of the start address
63 * @offset: offset of mapping relative to pfn
0abdd7a8 64 * @map_err_type: track whether dma_mapping_error() was checked
7c65aa3c
RD
65 * @stack_len: number of backtrace entries in @stack_entries
66 * @stack_entries: stack of backtrace history
0abdd7a8 67 */
f2f45e5f
JR
68struct dma_debug_entry {
69 struct list_head list;
70 struct device *dev;
f2f45e5f
JR
71 u64 dev_addr;
72 u64 size;
d3694f30 73 int type;
f2f45e5f
JR
74 int direction;
75 int sg_call_ents;
76 int sg_mapped_ents;
d3694f30
ED
77 unsigned long pfn;
78 size_t offset;
6c9c6d63 79 enum map_err_types map_err_type;
6c132d1b 80#ifdef CONFIG_STACKTRACE
746017ed
TG
81 unsigned int stack_len;
82 unsigned long stack_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
6c132d1b 83#endif
d3694f30 84} ____cacheline_aligned_in_smp;
f2f45e5f 85
c6a21d0b
NH
86typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
87
30dfa90c
JR
88struct hash_bucket {
89 struct list_head list;
90 spinlock_t lock;
5e76f564 91};
30dfa90c
JR
92
93/* Hash list to save the allocated dma addresses */
94static struct hash_bucket dma_entry_hash[HASH_SIZE];
3b1e79ed
JR
95/* List of pre-allocated dma_debug_entry's */
96static LIST_HEAD(free_entries);
97/* Lock for the list above */
98static DEFINE_SPINLOCK(free_entries_lock);
99
100/* Global disable flag - will be set in case of an error */
621a5f7a 101static bool global_disable __read_mostly;
3b1e79ed 102
2ce8e7ed
FF
103/* Early initialization disable flag, set at the end of dma_debug_init */
104static bool dma_debug_initialized __read_mostly;
105
01ce18b3
FF
106static inline bool dma_debug_disabled(void)
107{
2ce8e7ed 108 return global_disable || !dma_debug_initialized;
01ce18b3
FF
109}
110
788dcfa6
JR
111/* Global error count */
112static u32 error_count;
113
114/* Global error show enable*/
115static u32 show_all_errors __read_mostly;
116/* Number of errors to show */
117static u32 show_num_errors = 1;
118
3b1e79ed
JR
119static u32 num_free_entries;
120static u32 min_free_entries;
e6a1a89d 121static u32 nr_total_entries;
30dfa90c 122
59d3daaf 123/* number of preallocated entries requested by kernel cmdline */
bcebe324 124static u32 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
59d3daaf 125
2e507d84
JR
126/* per-driver filter related state */
127
128#define NAME_MAX_LEN 64
129
130static char current_driver_name[NAME_MAX_LEN] __read_mostly;
131static struct device_driver *current_driver __read_mostly;
132
133static DEFINE_RWLOCK(driver_name_lock);
788dcfa6 134
6c9c6d63
SK
135static const char *const maperr2str[] = {
136 [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
137 [MAP_ERR_NOT_CHECKED] = "dma map error not checked",
138 [MAP_ERR_CHECKED] = "dma map error checked",
139};
140
9bb50ed7
GS
141static const char *type2name[] = {
142 [dma_debug_single] = "single",
36d91e85 143 [dma_debug_sg] = "scatter-gather",
9bb50ed7
GS
144 [dma_debug_coherent] = "coherent",
145 [dma_debug_resource] = "resource",
146};
2d62ece1 147
23efed6f
CH
148static const char *dir2name[] = {
149 [DMA_BIDIRECTIONAL] = "DMA_BIDIRECTIONAL",
150 [DMA_TO_DEVICE] = "DMA_TO_DEVICE",
151 [DMA_FROM_DEVICE] = "DMA_FROM_DEVICE",
152 [DMA_NONE] = "DMA_NONE",
153};
2d62ece1
JR
154
155/*
156 * The access to some variables in this macro is racy. We can't use atomic_t
157 * here because all these variables are exported to debugfs. Some of them even
158 * writeable. This is also the reason why a lock won't help much. But anyway,
159 * the races are no big deal. Here is why:
160 *
161 * error_count: the addition is racy, but the worst thing that can happen is
162 * that we don't count some errors
163 * show_num_errors: the subtraction is racy. Also no big deal because in
164 * worst case this will result in one warning more in the
165 * system log than the user configured. This variable is
166 * writeable via debugfs.
167 */
6c132d1b
DW
168static inline void dump_entry_trace(struct dma_debug_entry *entry)
169{
170#ifdef CONFIG_STACKTRACE
171 if (entry) {
fc65104c 172 pr_warn("Mapped at:\n");
746017ed 173 stack_trace_print(entry->stack_entries, entry->stack_len, 0);
6c132d1b
DW
174 }
175#endif
176}
177
2e507d84
JR
178static bool driver_filter(struct device *dev)
179{
0bf84128
JR
180 struct device_driver *drv;
181 unsigned long flags;
182 bool ret;
183
2e507d84
JR
184 /* driver filter off */
185 if (likely(!current_driver_name[0]))
186 return true;
187
188 /* driver filter on and initialized */
ec9c96ef 189 if (current_driver && dev && dev->driver == current_driver)
2e507d84
JR
190 return true;
191
ec9c96ef
KM
192 /* driver filter on, but we can't filter on a NULL device... */
193 if (!dev)
194 return false;
195
0bf84128
JR
196 if (current_driver || !current_driver_name[0])
197 return false;
2e507d84 198
0bf84128 199 /* driver filter on but not yet initialized */
f3ff9247 200 drv = dev->driver;
0bf84128
JR
201 if (!drv)
202 return false;
203
204 /* lock to protect against change of current_driver_name */
205 read_lock_irqsave(&driver_name_lock, flags);
206
207 ret = false;
208 if (drv->name &&
209 strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
210 current_driver = drv;
211 ret = true;
2e507d84
JR
212 }
213
0bf84128 214 read_unlock_irqrestore(&driver_name_lock, flags);
0bf84128
JR
215
216 return ret;
2e507d84
JR
217}
218
ec9c96ef
KM
219#define err_printk(dev, entry, format, arg...) do { \
220 error_count += 1; \
221 if (driver_filter(dev) && \
222 (show_all_errors || show_num_errors > 0)) { \
f737b095 223 WARN(1, pr_fmt("%s %s: ") format, \
ec9c96ef
KM
224 dev ? dev_driver_string(dev) : "NULL", \
225 dev ? dev_name(dev) : "NULL", ## arg); \
226 dump_entry_trace(entry); \
227 } \
228 if (!show_all_errors && show_num_errors > 0) \
229 show_num_errors -= 1; \
2d62ece1
JR
230 } while (0);
231
30dfa90c
JR
232/*
233 * Hash related functions
234 *
235 * Every DMA-API request is saved into a struct dma_debug_entry. To
236 * have quick access to these structs they are stored into a hash.
237 */
238static int hash_fn(struct dma_debug_entry *entry)
239{
240 /*
241 * Hash function is based on the dma address.
242 * We use bits 20-27 here as the index into the hash
243 */
244 return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
245}
246
247/*
248 * Request exclusive access to a hash bucket for a given dma_debug_entry.
249 */
250static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
251 unsigned long *flags)
d5dfc80f 252 __acquires(&dma_entry_hash[idx].lock)
30dfa90c
JR
253{
254 int idx = hash_fn(entry);
255 unsigned long __flags;
256
257 spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
258 *flags = __flags;
259 return &dma_entry_hash[idx];
260}
261
262/*
263 * Give up exclusive access to the hash bucket
264 */
265static void put_hash_bucket(struct hash_bucket *bucket,
50f579a2 266 unsigned long flags)
d5dfc80f 267 __releases(&bucket->lock)
30dfa90c 268{
50f579a2 269 spin_unlock_irqrestore(&bucket->lock, flags);
30dfa90c
JR
270}
271
c6a21d0b
NH
272static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
273{
91ec37cc 274 return ((a->dev_addr == b->dev_addr) &&
c6a21d0b
NH
275 (a->dev == b->dev)) ? true : false;
276}
277
278static bool containing_match(struct dma_debug_entry *a,
279 struct dma_debug_entry *b)
280{
281 if (a->dev != b->dev)
282 return false;
283
284 if ((b->dev_addr <= a->dev_addr) &&
285 ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
286 return true;
287
288 return false;
289}
290
30dfa90c
JR
291/*
292 * Search a given entry in the hash bucket list
293 */
c6a21d0b
NH
294static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
295 struct dma_debug_entry *ref,
296 match_fn match)
30dfa90c 297{
7caf6a49 298 struct dma_debug_entry *entry, *ret = NULL;
fe73fbe1 299 int matches = 0, match_lvl, last_lvl = -1;
30dfa90c
JR
300
301 list_for_each_entry(entry, &bucket->list, list) {
c6a21d0b 302 if (!match(ref, entry))
7caf6a49
JR
303 continue;
304
305 /*
306 * Some drivers map the same physical address multiple
307 * times. Without a hardware IOMMU this results in the
308 * same device addresses being put into the dma-debug
309 * hash multiple times too. This can result in false
af901ca1 310 * positives being reported. Therefore we implement a
7caf6a49
JR
311 * best-fit algorithm here which returns the entry from
312 * the hash which fits best to the reference value
313 * instead of the first-fit.
314 */
315 matches += 1;
316 match_lvl = 0;
e5e8c5b9
JR
317 entry->size == ref->size ? ++match_lvl : 0;
318 entry->type == ref->type ? ++match_lvl : 0;
319 entry->direction == ref->direction ? ++match_lvl : 0;
320 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
7caf6a49 321
e5e8c5b9 322 if (match_lvl == 4) {
7caf6a49 323 /* perfect-fit - return the result */
30dfa90c 324 return entry;
7caf6a49
JR
325 } else if (match_lvl > last_lvl) {
326 /*
327 * We found an entry that fits better then the
fe73fbe1 328 * previous one or it is the 1st match.
7caf6a49
JR
329 */
330 last_lvl = match_lvl;
331 ret = entry;
332 }
30dfa90c
JR
333 }
334
7caf6a49
JR
335 /*
336 * If we have multiple matches but no perfect-fit, just return
337 * NULL.
338 */
339 ret = (matches == 1) ? ret : NULL;
340
341 return ret;
30dfa90c
JR
342}
343
c6a21d0b
NH
344static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
345 struct dma_debug_entry *ref)
346{
347 return __hash_bucket_find(bucket, ref, exact_match);
348}
349
350static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
351 struct dma_debug_entry *ref,
352 unsigned long *flags)
353{
354
c6a21d0b 355 struct dma_debug_entry *entry, index = *ref;
2995b800 356 int limit = min(HASH_SIZE, (index.dev_addr >> HASH_FN_SHIFT) + 1);
c6a21d0b 357
2995b800 358 for (int i = 0; i < limit; i++) {
a7a2c02a 359 entry = __hash_bucket_find(*bucket, ref, containing_match);
c6a21d0b
NH
360
361 if (entry)
362 return entry;
363
364 /*
365 * Nothing found, go back a hash bucket
366 */
50f579a2 367 put_hash_bucket(*bucket, *flags);
c6a21d0b
NH
368 index.dev_addr -= (1 << HASH_FN_SHIFT);
369 *bucket = get_hash_bucket(&index, flags);
370 }
371
372 return NULL;
373}
374
30dfa90c
JR
375/*
376 * Add an entry to a hash bucket
377 */
378static void hash_bucket_add(struct hash_bucket *bucket,
379 struct dma_debug_entry *entry)
380{
381 list_add_tail(&entry->list, &bucket->list);
382}
383
384/*
385 * Remove entry from a hash bucket list
386 */
387static void hash_bucket_del(struct dma_debug_entry *entry)
388{
389 list_del(&entry->list);
390}
391
0abdd7a8
DW
392static unsigned long long phys_addr(struct dma_debug_entry *entry)
393{
0e74b34d
NS
394 if (entry->type == dma_debug_resource)
395 return __pfn_to_phys(entry->pfn) + entry->offset;
396
0abdd7a8
DW
397 return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
398}
399
0abdd7a8 400/*
3b7a6418
DW
401 * For each mapping (initial cacheline in the case of
402 * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
403 * scatterlist, or the cacheline specified in dma_map_single) insert
404 * into this tree using the cacheline as the key. At
0abdd7a8 405 * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry. If
3b7a6418 406 * the entry already exists at insertion time add a tag as a reference
0abdd7a8 407 * count for the overlapping mappings. For now, the overlap tracking
3b7a6418
DW
408 * just ensures that 'unmaps' balance 'maps' before marking the
409 * cacheline idle, but we should also be flagging overlaps as an API
410 * violation.
0abdd7a8
DW
411 *
412 * Memory usage is mostly constrained by the maximum number of available
413 * dma-debug entries in that we need a free dma_debug_entry before
3b7a6418
DW
414 * inserting into the tree. In the case of dma_map_page and
415 * dma_alloc_coherent there is only one dma_debug_entry and one
416 * dma_active_cacheline entry to track per event. dma_map_sg(), on the
417 * other hand, consumes a single dma_debug_entry, but inserts 'nents'
418 * entries into the tree.
0abdd7a8 419 */
84bc4f1d 420static RADIX_TREE(dma_active_cacheline, GFP_ATOMIC);
0abdd7a8 421static DEFINE_SPINLOCK(radix_lock);
3b7a6418
DW
422#define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
423#define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
424#define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
0abdd7a8 425
3b7a6418
DW
426static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
427{
428 return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
429 (entry->offset >> L1_CACHE_SHIFT);
430}
431
432static int active_cacheline_read_overlap(phys_addr_t cln)
0abdd7a8
DW
433{
434 int overlap = 0, i;
435
436 for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
3b7a6418 437 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
0abdd7a8
DW
438 overlap |= 1 << i;
439 return overlap;
440}
441
3b7a6418 442static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
0abdd7a8
DW
443{
444 int i;
445
3b7a6418 446 if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
59f2e7df 447 return overlap;
0abdd7a8
DW
448
449 for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
450 if (overlap & 1 << i)
3b7a6418 451 radix_tree_tag_set(&dma_active_cacheline, cln, i);
0abdd7a8 452 else
3b7a6418 453 radix_tree_tag_clear(&dma_active_cacheline, cln, i);
0abdd7a8
DW
454
455 return overlap;
456}
457
3b7a6418 458static void active_cacheline_inc_overlap(phys_addr_t cln)
0abdd7a8 459{
3b7a6418 460 int overlap = active_cacheline_read_overlap(cln);
0abdd7a8 461
3b7a6418 462 overlap = active_cacheline_set_overlap(cln, ++overlap);
0abdd7a8
DW
463
464 /* If we overflowed the overlap counter then we're potentially
5848dc5b 465 * leaking dma-mappings.
0abdd7a8 466 */
3b7a6418 467 WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
f737b095 468 pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
3b7a6418 469 ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
0abdd7a8
DW
470}
471
3b7a6418 472static int active_cacheline_dec_overlap(phys_addr_t cln)
0abdd7a8 473{
3b7a6418 474 int overlap = active_cacheline_read_overlap(cln);
0abdd7a8 475
3b7a6418 476 return active_cacheline_set_overlap(cln, --overlap);
0abdd7a8
DW
477}
478
3b7a6418 479static int active_cacheline_insert(struct dma_debug_entry *entry)
0abdd7a8 480{
3b7a6418 481 phys_addr_t cln = to_cacheline_number(entry);
0abdd7a8
DW
482 unsigned long flags;
483 int rc;
484
3b7a6418
DW
485 /* If the device is not writing memory then we don't have any
486 * concerns about the cpu consuming stale data. This mitigates
487 * legitimate usages of overlapping mappings.
488 */
489 if (entry->direction == DMA_TO_DEVICE)
490 return 0;
491
0abdd7a8 492 spin_lock_irqsave(&radix_lock, flags);
3b7a6418 493 rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
0abdd7a8 494 if (rc == -EEXIST)
3b7a6418 495 active_cacheline_inc_overlap(cln);
0abdd7a8
DW
496 spin_unlock_irqrestore(&radix_lock, flags);
497
498 return rc;
499}
500
3b7a6418 501static void active_cacheline_remove(struct dma_debug_entry *entry)
0abdd7a8 502{
3b7a6418 503 phys_addr_t cln = to_cacheline_number(entry);
0abdd7a8
DW
504 unsigned long flags;
505
3b7a6418
DW
506 /* ...mirror the insert case */
507 if (entry->direction == DMA_TO_DEVICE)
508 return;
509
0abdd7a8 510 spin_lock_irqsave(&radix_lock, flags);
59f2e7df 511 /* since we are counting overlaps the final put of the
3b7a6418
DW
512 * cacheline will occur when the overlap count is 0.
513 * active_cacheline_dec_overlap() returns -1 in that case
59f2e7df 514 */
3b7a6418
DW
515 if (active_cacheline_dec_overlap(cln) < 0)
516 radix_tree_delete(&dma_active_cacheline, cln);
0abdd7a8
DW
517 spin_unlock_irqrestore(&radix_lock, flags);
518}
519
bd89d69a
DN
520/*
521 * Dump mappings entries on kernel space for debugging purposes
522 */
523void debug_dma_dump_mappings(struct device *dev)
524{
525 int idx;
526 phys_addr_t cln;
527
528 for (idx = 0; idx < HASH_SIZE; idx++) {
529 struct hash_bucket *bucket = &dma_entry_hash[idx];
530 struct dma_debug_entry *entry;
531 unsigned long flags;
532
533 spin_lock_irqsave(&bucket->lock, flags);
534 list_for_each_entry(entry, &bucket->list, list) {
535 if (!dev || dev == entry->dev) {
536 cln = to_cacheline_number(entry);
537 dev_info(entry->dev,
b31507dc 538 "%s idx %d P=%llx N=%lx D=%llx L=%llx cln=%pa %s %s\n",
bd89d69a
DN
539 type2name[entry->type], idx,
540 phys_addr(entry), entry->pfn,
541 entry->dev_addr, entry->size,
b31507dc 542 &cln, dir2name[entry->direction],
bd89d69a
DN
543 maperr2str[entry->map_err_type]);
544 }
545 }
546 spin_unlock_irqrestore(&bucket->lock, flags);
547
548 cond_resched();
549 }
550}
551
552/*
553 * Dump mappings entries on user space via debugfs
554 */
555static int dump_show(struct seq_file *seq, void *v)
556{
557 int idx;
558 phys_addr_t cln;
559
560 for (idx = 0; idx < HASH_SIZE; idx++) {
561 struct hash_bucket *bucket = &dma_entry_hash[idx];
562 struct dma_debug_entry *entry;
563 unsigned long flags;
564
565 spin_lock_irqsave(&bucket->lock, flags);
566 list_for_each_entry(entry, &bucket->list, list) {
567 cln = to_cacheline_number(entry);
568 seq_printf(seq,
b31507dc 569 "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx cln=%pa %s %s\n",
bd89d69a
DN
570 dev_driver_string(entry->dev),
571 dev_name(entry->dev),
572 type2name[entry->type], idx,
573 phys_addr(entry), entry->pfn,
574 entry->dev_addr, entry->size,
b31507dc 575 &cln, dir2name[entry->direction],
bd89d69a
DN
576 maperr2str[entry->map_err_type]);
577 }
578 spin_unlock_irqrestore(&bucket->lock, flags);
579 }
580 return 0;
581}
582DEFINE_SHOW_ATTRIBUTE(dump);
583
30dfa90c
JR
584/*
585 * Wrapper function for adding an entry to the hash.
586 * This function takes care of locking itself.
587 */
c2bbf9d1 588static void add_dma_entry(struct dma_debug_entry *entry, unsigned long attrs)
30dfa90c
JR
589{
590 struct hash_bucket *bucket;
591 unsigned long flags;
0abdd7a8 592 int rc;
30dfa90c
JR
593
594 bucket = get_hash_bucket(entry, &flags);
595 hash_bucket_add(bucket, entry);
50f579a2 596 put_hash_bucket(bucket, flags);
0abdd7a8 597
3b7a6418 598 rc = active_cacheline_insert(entry);
0abdd7a8 599 if (rc == -ENOMEM) {
e19f8fa6 600 pr_err_once("cacheline tracking ENOMEM, dma-debug disabled\n");
0abdd7a8 601 global_disable = true;
c2bbf9d1 602 } else if (rc == -EEXIST && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
510e1a72
HM
603 err_printk(entry->dev, entry,
604 "cacheline tracking EEXIST, overlapping mappings aren't supported\n");
0abdd7a8 605 }
30dfa90c
JR
606}
607
ad78dee0 608static int dma_debug_create_entries(gfp_t gfp)
2b9d9ac0 609{
ad78dee0 610 struct dma_debug_entry *entry;
2b9d9ac0
RM
611 int i;
612
ad78dee0
RM
613 entry = (void *)get_zeroed_page(gfp);
614 if (!entry)
615 return -ENOMEM;
2b9d9ac0 616
ad78dee0
RM
617 for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
618 list_add_tail(&entry[i].list, &free_entries);
2b9d9ac0 619
ad78dee0
RM
620 num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
621 nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
2b9d9ac0
RM
622
623 return 0;
2b9d9ac0
RM
624}
625
e6a1a89d
FT
626static struct dma_debug_entry *__dma_entry_alloc(void)
627{
628 struct dma_debug_entry *entry;
629
630 entry = list_entry(free_entries.next, struct dma_debug_entry, list);
631 list_del(&entry->list);
632 memset(entry, 0, sizeof(*entry));
633
634 num_free_entries -= 1;
635 if (num_free_entries < min_free_entries)
636 min_free_entries = num_free_entries;
637
638 return entry;
639}
640
fb5a4315
SS
641/*
642 * This should be called outside of free_entries_lock scope to avoid potential
643 * deadlocks with serial consoles that use DMA.
644 */
645static void __dma_entry_alloc_check_leak(u32 nr_entries)
ceb51173 646{
fb5a4315 647 u32 tmp = nr_entries % nr_prealloc_entries;
ceb51173
RM
648
649 /* Shout each time we tick over some multiple of the initial pool */
650 if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
651 pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
fb5a4315
SS
652 nr_entries,
653 (nr_entries / nr_prealloc_entries));
ceb51173
RM
654 }
655}
656
3b1e79ed
JR
657/* struct dma_entry allocator
658 *
659 * The next two functions implement the allocator for
660 * struct dma_debug_entries.
661 */
662static struct dma_debug_entry *dma_entry_alloc(void)
663{
fb5a4315 664 bool alloc_check_leak = false;
29cdd4e4 665 struct dma_debug_entry *entry;
3b1e79ed 666 unsigned long flags;
fb5a4315 667 u32 nr_entries;
3b1e79ed
JR
668
669 spin_lock_irqsave(&free_entries_lock, flags);
2b9d9ac0 670 if (num_free_entries == 0) {
ad78dee0 671 if (dma_debug_create_entries(GFP_ATOMIC)) {
2b9d9ac0
RM
672 global_disable = true;
673 spin_unlock_irqrestore(&free_entries_lock, flags);
674 pr_err("debugging out of memory - disabling\n");
675 return NULL;
676 }
fb5a4315
SS
677 alloc_check_leak = true;
678 nr_entries = nr_total_entries;
3b1e79ed
JR
679 }
680
e6a1a89d 681 entry = __dma_entry_alloc();
3b1e79ed 682
29cdd4e4
JK
683 spin_unlock_irqrestore(&free_entries_lock, flags);
684
fb5a4315
SS
685 if (alloc_check_leak)
686 __dma_entry_alloc_check_leak(nr_entries);
687
6c132d1b 688#ifdef CONFIG_STACKTRACE
746017ed
TG
689 entry->stack_len = stack_trace_save(entry->stack_entries,
690 ARRAY_SIZE(entry->stack_entries),
691 1);
6c132d1b 692#endif
3b1e79ed
JR
693 return entry;
694}
695
696static void dma_entry_free(struct dma_debug_entry *entry)
697{
698 unsigned long flags;
699
3b7a6418 700 active_cacheline_remove(entry);
0abdd7a8 701
3b1e79ed
JR
702 /*
703 * add to beginning of the list - this way the entries are
704 * more likely cache hot when they are reallocated.
705 */
706 spin_lock_irqsave(&free_entries_lock, flags);
707 list_add(&entry->list, &free_entries);
708 num_free_entries += 1;
709 spin_unlock_irqrestore(&free_entries_lock, flags);
710}
711
6bf07871
JR
712/*
713 * DMA-API debugging init code
714 *
715 * The init code does two things:
716 * 1. Initialize core data structures
717 * 2. Preallocate a given number of dma_debug_entry structs
718 */
719
8a6fc708
JR
720static ssize_t filter_read(struct file *file, char __user *user_buf,
721 size_t count, loff_t *ppos)
722{
8a6fc708 723 char buf[NAME_MAX_LEN + 1];
c17e2cf7 724 unsigned long flags;
8a6fc708
JR
725 int len;
726
727 if (!current_driver_name[0])
728 return 0;
729
730 /*
731 * We can't copy to userspace directly because current_driver_name can
732 * only be read under the driver_name_lock with irqs disabled. So
733 * create a temporary copy first.
734 */
735 read_lock_irqsave(&driver_name_lock, flags);
736 len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
737 read_unlock_irqrestore(&driver_name_lock, flags);
738
739 return simple_read_from_buffer(user_buf, count, ppos, buf, len);
740}
741
742static ssize_t filter_write(struct file *file, const char __user *userbuf,
743 size_t count, loff_t *ppos)
744{
8a6fc708 745 char buf[NAME_MAX_LEN];
c17e2cf7
JR
746 unsigned long flags;
747 size_t len;
8a6fc708
JR
748 int i;
749
750 /*
751 * We can't copy from userspace directly. Access to
752 * current_driver_name is protected with a write_lock with irqs
753 * disabled. Since copy_from_user can fault and may sleep we
754 * need to copy to temporary buffer first
755 */
e7ed70ee 756 len = min(count, (size_t)(NAME_MAX_LEN - 1));
8a6fc708
JR
757 if (copy_from_user(buf, userbuf, len))
758 return -EFAULT;
759
760 buf[len] = 0;
761
762 write_lock_irqsave(&driver_name_lock, flags);
763
31232509
JR
764 /*
765 * Now handle the string we got from userspace very carefully.
8a6fc708
JR
766 * The rules are:
767 * - only use the first token we got
768 * - token delimiter is everything looking like a space
769 * character (' ', '\n', '\t' ...)
770 *
771 */
772 if (!isalnum(buf[0])) {
773 /*
31232509 774 * If the first character userspace gave us is not
8a6fc708
JR
775 * alphanumerical then assume the filter should be
776 * switched off.
777 */
778 if (current_driver_name[0])
f737b095 779 pr_info("switching off dma-debug driver filter\n");
8a6fc708
JR
780 current_driver_name[0] = 0;
781 current_driver = NULL;
782 goto out_unlock;
783 }
784
785 /*
786 * Now parse out the first token and use it as the name for the
787 * driver to filter for.
788 */
39a37ce1 789 for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
8a6fc708
JR
790 current_driver_name[i] = buf[i];
791 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
792 break;
793 }
794 current_driver_name[i] = 0;
795 current_driver = NULL;
796
f737b095 797 pr_info("enable driver filter for driver [%s]\n",
e7ed70ee 798 current_driver_name);
8a6fc708
JR
799
800out_unlock:
801 write_unlock_irqrestore(&driver_name_lock, flags);
802
803 return count;
804}
805
aeb583d0 806static const struct file_operations filter_fops = {
8a6fc708
JR
807 .read = filter_read,
808 .write = filter_write,
6038f373 809 .llseek = default_llseek,
8a6fc708
JR
810};
811
173735c3 812static int __init dma_debug_fs_init(void)
788dcfa6 813{
8e4d81b9 814 struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
788dcfa6 815
8e4d81b9
GKH
816 debugfs_create_bool("disabled", 0444, dentry, &global_disable);
817 debugfs_create_u32("error_count", 0444, dentry, &error_count);
818 debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
819 debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
820 debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
821 debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
822 debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
823 debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
0a3b192c 824 debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
173735c3
AI
825
826 return 0;
788dcfa6 827}
173735c3 828core_initcall_sync(dma_debug_fs_init);
788dcfa6 829
ba4b87ad 830static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
ed888aef
JR
831{
832 struct dma_debug_entry *entry;
833 unsigned long flags;
834 int count = 0, i;
835
836 for (i = 0; i < HASH_SIZE; ++i) {
6a5cd60b 837 spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
ed888aef 838 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
ba4b87ad 839 if (entry->dev == dev) {
ed888aef 840 count += 1;
ba4b87ad
SG
841 *out_entry = entry;
842 }
ed888aef 843 }
6a5cd60b 844 spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
ed888aef
JR
845 }
846
847 return count;
848}
849
a8fe9ea2 850static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
ed888aef
JR
851{
852 struct device *dev = data;
3f649ab7 853 struct dma_debug_entry *entry;
ed888aef
JR
854 int count;
855
01ce18b3 856 if (dma_debug_disabled())
a8fe9ea2 857 return 0;
ed888aef
JR
858
859 switch (action) {
860 case BUS_NOTIFY_UNBOUND_DRIVER:
ba4b87ad 861 count = device_dma_allocations(dev, &entry);
ed888aef
JR
862 if (count == 0)
863 break;
f737b095 864 err_printk(dev, entry, "device driver has pending "
ed888aef 865 "DMA allocations while released from device "
ba4b87ad
SG
866 "[count=%d]\n"
867 "One of leaked entries details: "
868 "[device address=0x%016llx] [size=%llu bytes] "
869 "[mapped with %s] [mapped as %s]\n",
870 count, entry->dev_addr, entry->size,
871 dir2name[entry->direction], type2name[entry->type]);
ed888aef
JR
872 break;
873 default:
874 break;
875 }
876
877 return 0;
878}
879
86438841 880void dma_debug_add_bus(const struct bus_type *bus)
41531c8f 881{
ed888aef
JR
882 struct notifier_block *nb;
883
01ce18b3 884 if (dma_debug_disabled())
f797d988
SR
885 return;
886
ed888aef
JR
887 nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
888 if (nb == NULL) {
e7ed70ee 889 pr_err("dma_debug_add_bus: out of memory\n");
ed888aef
JR
890 return;
891 }
892
893 nb->notifier_call = dma_debug_device_change;
894
895 bus_register_notifier(bus, nb);
41531c8f 896}
788dcfa6 897
15b28bbc 898static int dma_debug_init(void)
6bf07871 899{
ad78dee0 900 int i, nr_pages;
6bf07871 901
2ce8e7ed
FF
902 /* Do not use dma_debug_initialized here, since we really want to be
903 * called to set dma_debug_initialized
904 */
905 if (global_disable)
15b28bbc 906 return 0;
6bf07871
JR
907
908 for (i = 0; i < HASH_SIZE; ++i) {
909 INIT_LIST_HEAD(&dma_entry_hash[i].list);
b0a5b83e 910 spin_lock_init(&dma_entry_hash[i].lock);
6bf07871
JR
911 }
912
ad78dee0
RM
913 nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
914 for (i = 0; i < nr_pages; ++i)
915 dma_debug_create_entries(GFP_KERNEL);
916 if (num_free_entries >= nr_prealloc_entries) {
917 pr_info("preallocated %d debug entries\n", nr_total_entries);
918 } else if (num_free_entries > 0) {
919 pr_warn("%d debug entries requested but only %d allocated\n",
920 nr_prealloc_entries, nr_total_entries);
921 } else {
f737b095 922 pr_err("debugging out of memory error - disabled\n");
6bf07871
JR
923 global_disable = true;
924
15b28bbc 925 return 0;
6bf07871 926 }
2b9d9ac0 927 min_free_entries = num_free_entries;
e6a1a89d 928
2ce8e7ed
FF
929 dma_debug_initialized = true;
930
f737b095 931 pr_info("debugging enabled by kernel config\n");
15b28bbc 932 return 0;
6bf07871 933}
15b28bbc 934core_initcall(dma_debug_init);
6bf07871 935
59d3daaf
JR
936static __init int dma_debug_cmdline(char *str)
937{
938 if (!str)
939 return -EINVAL;
940
941 if (strncmp(str, "off", 3) == 0) {
f737b095 942 pr_info("debugging disabled on kernel command line\n");
59d3daaf
JR
943 global_disable = true;
944 }
945
80e43909 946 return 1;
59d3daaf
JR
947}
948
949static __init int dma_debug_entries_cmdline(char *str)
950{
59d3daaf
JR
951 if (!str)
952 return -EINVAL;
bcebe324
CH
953 if (!get_option(&str, &nr_prealloc_entries))
954 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
80e43909 955 return 1;
59d3daaf
JR
956}
957
958__setup("dma_debug=", dma_debug_cmdline);
959__setup("dma_debug_entries=", dma_debug_entries_cmdline);
960
2d62ece1
JR
961static void check_unmap(struct dma_debug_entry *ref)
962{
963 struct dma_debug_entry *entry;
964 struct hash_bucket *bucket;
965 unsigned long flags;
966
2d62ece1 967 bucket = get_hash_bucket(ref, &flags);
c6a21d0b 968 entry = bucket_find_exact(bucket, ref);
2d62ece1
JR
969
970 if (!entry) {
8d640a51 971 /* must drop lock before calling dma_mapping_error */
50f579a2 972 put_hash_bucket(bucket, flags);
8d640a51 973
bfe0fb0f
SK
974 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
975 err_printk(ref->dev, NULL,
f737b095 976 "device driver tries to free an "
8d640a51
AD
977 "invalid DMA memory address\n");
978 } else {
979 err_printk(ref->dev, NULL,
f737b095 980 "device driver tries to free DMA "
8d640a51
AD
981 "memory it has not allocated [device "
982 "address=0x%016llx] [size=%llu bytes]\n",
983 ref->dev_addr, ref->size);
bfe0fb0f 984 }
8d640a51 985 return;
2d62ece1
JR
986 }
987
988 if (ref->size != entry->size) {
f737b095 989 err_printk(ref->dev, entry, "device driver frees "
2d62ece1
JR
990 "DMA memory with different size "
991 "[device address=0x%016llx] [map size=%llu bytes] "
992 "[unmap size=%llu bytes]\n",
993 ref->dev_addr, entry->size, ref->size);
994 }
995
996 if (ref->type != entry->type) {
f737b095 997 err_printk(ref->dev, entry, "device driver frees "
2d62ece1
JR
998 "DMA memory with wrong function "
999 "[device address=0x%016llx] [size=%llu bytes] "
1000 "[mapped as %s] [unmapped as %s]\n",
1001 ref->dev_addr, ref->size,
1002 type2name[entry->type], type2name[ref->type]);
1003 } else if ((entry->type == dma_debug_coherent) &&
0abdd7a8 1004 (phys_addr(ref) != phys_addr(entry))) {
f737b095 1005 err_printk(ref->dev, entry, "device driver frees "
2d62ece1
JR
1006 "DMA memory with different CPU address "
1007 "[device address=0x%016llx] [size=%llu bytes] "
59a40e70
JR
1008 "[cpu alloc address=0x%016llx] "
1009 "[cpu free address=0x%016llx]",
2d62ece1 1010 ref->dev_addr, ref->size,
0abdd7a8
DW
1011 phys_addr(entry),
1012 phys_addr(ref));
2d62ece1
JR
1013 }
1014
1015 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1016 ref->sg_call_ents != entry->sg_call_ents) {
f737b095 1017 err_printk(ref->dev, entry, "device driver frees "
2d62ece1
JR
1018 "DMA sg list with different entry count "
1019 "[map count=%d] [unmap count=%d]\n",
1020 entry->sg_call_ents, ref->sg_call_ents);
1021 }
1022
1023 /*
1024 * This may be no bug in reality - but most implementations of the
1025 * DMA API don't handle this properly, so check for it here
1026 */
1027 if (ref->direction != entry->direction) {
f737b095 1028 err_printk(ref->dev, entry, "device driver frees "
2d62ece1
JR
1029 "DMA memory with different direction "
1030 "[device address=0x%016llx] [size=%llu bytes] "
1031 "[mapped with %s] [unmapped with %s]\n",
1032 ref->dev_addr, ref->size,
1033 dir2name[entry->direction],
1034 dir2name[ref->direction]);
1035 }
1036
a5759b2b
MC
1037 /*
1038 * Drivers should use dma_mapping_error() to check the returned
1039 * addresses of dma_map_single() and dma_map_page().
985098a0 1040 * If not, print this warning message. See Documentation/core-api/dma-api.rst.
a5759b2b 1041 */
6c9c6d63
SK
1042 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1043 err_printk(ref->dev, entry,
f737b095 1044 "device driver failed to check map error"
6c9c6d63
SK
1045 "[device address=0x%016llx] [size=%llu bytes] "
1046 "[mapped as %s]",
1047 ref->dev_addr, ref->size,
1048 type2name[entry->type]);
1049 }
1050
2d62ece1
JR
1051 hash_bucket_del(entry);
1052 dma_entry_free(entry);
1053
50f579a2 1054 put_hash_bucket(bucket, flags);
2d62ece1
JR
1055}
1056
b4a0f533
AL
1057static void check_for_stack(struct device *dev,
1058 struct page *page, size_t offset)
2d62ece1 1059{
b4a0f533
AL
1060 void *addr;
1061 struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1062
1063 if (!stack_vm_area) {
1064 /* Stack is direct-mapped. */
1065 if (PageHighMem(page))
1066 return;
1067 addr = page_address(page) + offset;
1068 if (object_is_on_stack(addr))
f737b095 1069 err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
b4a0f533
AL
1070 } else {
1071 /* Stack is vmalloced. */
1072 int i;
1073
1074 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1075 if (page != stack_vm_area->pages[i])
1076 continue;
1077
1078 addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
f737b095 1079 err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
b4a0f533
AL
1080 break;
1081 }
1082 }
2d62ece1
JR
1083}
1084
f39d1b97 1085static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
2e34bde1 1086{
1d7db834
KW
1087 if (memory_intersects(_stext, _etext, addr, len) ||
1088 memory_intersects(__start_rodata, __end_rodata, addr, len))
f737b095 1089 err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
2e34bde1
JR
1090}
1091
aa010efb
JR
1092static void check_sync(struct device *dev,
1093 struct dma_debug_entry *ref,
1094 bool to_cpu)
2d62ece1 1095{
2d62ece1
JR
1096 struct dma_debug_entry *entry;
1097 struct hash_bucket *bucket;
1098 unsigned long flags;
1099
aa010efb 1100 bucket = get_hash_bucket(ref, &flags);
2d62ece1 1101
c6a21d0b 1102 entry = bucket_find_contain(&bucket, ref, &flags);
2d62ece1
JR
1103
1104 if (!entry) {
f737b095 1105 err_printk(dev, NULL, "device driver tries "
2d62ece1
JR
1106 "to sync DMA memory it has not allocated "
1107 "[device address=0x%016llx] [size=%llu bytes]\n",
aa010efb 1108 (unsigned long long)ref->dev_addr, ref->size);
2d62ece1
JR
1109 goto out;
1110 }
1111
aa010efb 1112 if (ref->size > entry->size) {
f737b095 1113 err_printk(dev, entry, "device driver syncs"
2d62ece1
JR
1114 " DMA memory outside allocated range "
1115 "[device address=0x%016llx] "
aa010efb
JR
1116 "[allocation size=%llu bytes] "
1117 "[sync offset+size=%llu]\n",
1118 entry->dev_addr, entry->size,
1119 ref->size);
2d62ece1
JR
1120 }
1121
42d53b4f
KH
1122 if (entry->direction == DMA_BIDIRECTIONAL)
1123 goto out;
1124
aa010efb 1125 if (ref->direction != entry->direction) {
f737b095 1126 err_printk(dev, entry, "device driver syncs "
2d62ece1
JR
1127 "DMA memory with different direction "
1128 "[device address=0x%016llx] [size=%llu bytes] "
1129 "[mapped with %s] [synced with %s]\n",
aa010efb 1130 (unsigned long long)ref->dev_addr, entry->size,
2d62ece1 1131 dir2name[entry->direction],
aa010efb 1132 dir2name[ref->direction]);
2d62ece1
JR
1133 }
1134
2d62ece1 1135 if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
aa010efb 1136 !(ref->direction == DMA_TO_DEVICE))
f737b095 1137 err_printk(dev, entry, "device driver syncs "
2d62ece1
JR
1138 "device read-only DMA memory for cpu "
1139 "[device address=0x%016llx] [size=%llu bytes] "
1140 "[mapped with %s] [synced with %s]\n",
aa010efb 1141 (unsigned long long)ref->dev_addr, entry->size,
2d62ece1 1142 dir2name[entry->direction],
aa010efb 1143 dir2name[ref->direction]);
2d62ece1
JR
1144
1145 if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
aa010efb 1146 !(ref->direction == DMA_FROM_DEVICE))
f737b095 1147 err_printk(dev, entry, "device driver syncs "
2d62ece1
JR
1148 "device write-only DMA memory to device "
1149 "[device address=0x%016llx] [size=%llu bytes] "
1150 "[mapped with %s] [synced with %s]\n",
aa010efb 1151 (unsigned long long)ref->dev_addr, entry->size,
2d62ece1 1152 dir2name[entry->direction],
aa010efb 1153 dir2name[ref->direction]);
2d62ece1 1154
7f830642
RM
1155 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1156 ref->sg_call_ents != entry->sg_call_ents) {
f737b095 1157 err_printk(ref->dev, entry, "device driver syncs "
7f830642
RM
1158 "DMA sg list with different entry count "
1159 "[map count=%d] [sync count=%d]\n",
1160 entry->sg_call_ents, ref->sg_call_ents);
1161 }
1162
2d62ece1 1163out:
50f579a2 1164 put_hash_bucket(bucket, flags);
2d62ece1
JR
1165}
1166
78c47830
RM
1167static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1168{
1169#ifdef CONFIG_DMA_API_DEBUG_SG
1170 unsigned int max_seg = dma_get_max_seg_size(dev);
1171 u64 start, end, boundary = dma_get_seg_boundary(dev);
1172
1173 /*
1174 * Either the driver forgot to set dma_parms appropriately, or
1175 * whoever generated the list forgot to check them.
1176 */
1177 if (sg->length > max_seg)
f737b095 1178 err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
78c47830
RM
1179 sg->length, max_seg);
1180 /*
1181 * In some cases this could potentially be the DMA API
1182 * implementation's fault, but it would usually imply that
1183 * the scatterlist was built inappropriately to begin with.
1184 */
1185 start = sg_dma_address(sg);
1186 end = start + sg_dma_len(sg) - 1;
1187 if ((start ^ end) & ~boundary)
f737b095 1188 err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
78c47830
RM
1189 start, end, boundary);
1190#endif
1191}
1192
99c65fa7
SB
1193void debug_dma_map_single(struct device *dev, const void *addr,
1194 unsigned long len)
1195{
1196 if (unlikely(dma_debug_disabled()))
1197 return;
1198
1199 if (!virt_addr_valid(addr))
f737b095 1200 err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
99c65fa7
SB
1201 addr, len);
1202
1203 if (is_vmalloc_addr(addr))
f737b095 1204 err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
99c65fa7
SB
1205 addr, len);
1206}
1207EXPORT_SYMBOL(debug_dma_map_single);
1208
f62bc980 1209void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
c2bbf9d1
HM
1210 size_t size, int direction, dma_addr_t dma_addr,
1211 unsigned long attrs)
f62bc980
JR
1212{
1213 struct dma_debug_entry *entry;
1214
01ce18b3 1215 if (unlikely(dma_debug_disabled()))
f62bc980
JR
1216 return;
1217
bfe0fb0f 1218 if (dma_mapping_error(dev, dma_addr))
f62bc980
JR
1219 return;
1220
1221 entry = dma_entry_alloc();
1222 if (!entry)
1223 return;
1224
1225 entry->dev = dev;
2e05ea5c 1226 entry->type = dma_debug_single;
0abdd7a8 1227 entry->pfn = page_to_pfn(page);
a97740f8 1228 entry->offset = offset;
f62bc980
JR
1229 entry->dev_addr = dma_addr;
1230 entry->size = size;
1231 entry->direction = direction;
6c9c6d63 1232 entry->map_err_type = MAP_ERR_NOT_CHECKED;
f62bc980 1233
b4a0f533
AL
1234 check_for_stack(dev, page, offset);
1235
9537a48e 1236 if (!PageHighMem(page)) {
f39d1b97
IM
1237 void *addr = page_address(page) + offset;
1238
2e34bde1 1239 check_for_illegal_area(dev, addr, size);
f62bc980
JR
1240 }
1241
c2bbf9d1 1242 add_dma_entry(entry, attrs);
f62bc980 1243}
f62bc980 1244
6c9c6d63
SK
1245void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1246{
1247 struct dma_debug_entry ref;
1248 struct dma_debug_entry *entry;
1249 struct hash_bucket *bucket;
1250 unsigned long flags;
1251
01ce18b3 1252 if (unlikely(dma_debug_disabled()))
6c9c6d63
SK
1253 return;
1254
1255 ref.dev = dev;
1256 ref.dev_addr = dma_addr;
1257 bucket = get_hash_bucket(&ref, &flags);
6c9c6d63 1258
96e7d7a1
AD
1259 list_for_each_entry(entry, &bucket->list, list) {
1260 if (!exact_match(&ref, entry))
1261 continue;
1262
1263 /*
1264 * The same physical address can be mapped multiple
1265 * times. Without a hardware IOMMU this results in the
1266 * same device addresses being put into the dma-debug
1267 * hash multiple times too. This can result in false
1268 * positives being reported. Therefore we implement a
1269 * best-fit algorithm here which updates the first entry
1270 * from the hash which fits the reference value and is
1271 * not currently listed as being checked.
1272 */
1273 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1274 entry->map_err_type = MAP_ERR_CHECKED;
1275 break;
1276 }
1277 }
6c9c6d63 1278
50f579a2 1279 put_hash_bucket(bucket, flags);
6c9c6d63
SK
1280}
1281EXPORT_SYMBOL(debug_dma_mapping_error);
1282
479623fd 1283void debug_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
2e05ea5c 1284 size_t size, int direction)
f62bc980
JR
1285{
1286 struct dma_debug_entry ref = {
2e05ea5c 1287 .type = dma_debug_single,
f62bc980 1288 .dev = dev,
479623fd 1289 .dev_addr = dma_addr,
f62bc980
JR
1290 .size = size,
1291 .direction = direction,
1292 };
1293
01ce18b3 1294 if (unlikely(dma_debug_disabled()))
f62bc980 1295 return;
f62bc980
JR
1296 check_unmap(&ref);
1297}
f62bc980 1298
972aa45c 1299void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
c2bbf9d1
HM
1300 int nents, int mapped_ents, int direction,
1301 unsigned long attrs)
972aa45c
JR
1302{
1303 struct dma_debug_entry *entry;
1304 struct scatterlist *s;
1305 int i;
1306
01ce18b3 1307 if (unlikely(dma_debug_disabled()))
972aa45c
JR
1308 return;
1309
293d92cb
GS
1310 for_each_sg(sg, s, nents, i) {
1311 check_for_stack(dev, sg_page(s), s->offset);
1312 if (!PageHighMem(sg_page(s)))
1313 check_for_illegal_area(dev, sg_virt(s), s->length);
1314 }
1315
972aa45c
JR
1316 for_each_sg(sg, s, mapped_ents, i) {
1317 entry = dma_entry_alloc();
1318 if (!entry)
1319 return;
1320
1321 entry->type = dma_debug_sg;
1322 entry->dev = dev;
0abdd7a8 1323 entry->pfn = page_to_pfn(sg_page(s));
a97740f8 1324 entry->offset = s->offset;
884d0597 1325 entry->size = sg_dma_len(s);
15aedea4 1326 entry->dev_addr = sg_dma_address(s);
972aa45c
JR
1327 entry->direction = direction;
1328 entry->sg_call_ents = nents;
1329 entry->sg_mapped_ents = mapped_ents;
1330
78c47830
RM
1331 check_sg_segment(dev, s);
1332
c2bbf9d1 1333 add_dma_entry(entry, attrs);
972aa45c
JR
1334 }
1335}
972aa45c 1336
aa010efb
JR
1337static int get_nr_mapped_entries(struct device *dev,
1338 struct dma_debug_entry *ref)
88f3907f 1339{
aa010efb 1340 struct dma_debug_entry *entry;
88f3907f
FT
1341 struct hash_bucket *bucket;
1342 unsigned long flags;
c17e2cf7 1343 int mapped_ents;
88f3907f 1344
aa010efb 1345 bucket = get_hash_bucket(ref, &flags);
c6a21d0b 1346 entry = bucket_find_exact(bucket, ref);
c17e2cf7 1347 mapped_ents = 0;
88f3907f 1348
88f3907f
FT
1349 if (entry)
1350 mapped_ents = entry->sg_mapped_ents;
50f579a2 1351 put_hash_bucket(bucket, flags);
88f3907f
FT
1352
1353 return mapped_ents;
1354}
1355
972aa45c
JR
1356void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1357 int nelems, int dir)
1358{
972aa45c
JR
1359 struct scatterlist *s;
1360 int mapped_ents = 0, i;
972aa45c 1361
01ce18b3 1362 if (unlikely(dma_debug_disabled()))
972aa45c
JR
1363 return;
1364
1365 for_each_sg(sglist, s, nelems, i) {
1366
1367 struct dma_debug_entry ref = {
1368 .type = dma_debug_sg,
1369 .dev = dev,
0abdd7a8
DW
1370 .pfn = page_to_pfn(sg_page(s)),
1371 .offset = s->offset,
15aedea4 1372 .dev_addr = sg_dma_address(s),
884d0597 1373 .size = sg_dma_len(s),
972aa45c 1374 .direction = dir,
e5e8c5b9 1375 .sg_call_ents = nelems,
972aa45c
JR
1376 };
1377
1378 if (mapped_ents && i >= mapped_ents)
1379 break;
1380
e5e8c5b9 1381 if (!i)
aa010efb 1382 mapped_ents = get_nr_mapped_entries(dev, &ref);
972aa45c
JR
1383
1384 check_unmap(&ref);
1385 }
1386}
972aa45c 1387
6bfd4498 1388void debug_dma_alloc_coherent(struct device *dev, size_t size,
c2bbf9d1
HM
1389 dma_addr_t dma_addr, void *virt,
1390 unsigned long attrs)
6bfd4498
JR
1391{
1392 struct dma_debug_entry *entry;
1393
01ce18b3 1394 if (unlikely(dma_debug_disabled()))
6bfd4498
JR
1395 return;
1396
1397 if (unlikely(virt == NULL))
1398 return;
1399
af1da686
MC
1400 /* handle vmalloc and linear addresses */
1401 if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
6bfd4498
JR
1402 return;
1403
af1da686
MC
1404 entry = dma_entry_alloc();
1405 if (!entry)
3aaabbf1
MC
1406 return;
1407
6bfd4498
JR
1408 entry->type = dma_debug_coherent;
1409 entry->dev = dev;
e57d0552 1410 entry->offset = offset_in_page(virt);
6bfd4498
JR
1411 entry->size = size;
1412 entry->dev_addr = dma_addr;
1413 entry->direction = DMA_BIDIRECTIONAL;
1414
3aaabbf1
MC
1415 if (is_vmalloc_addr(virt))
1416 entry->pfn = vmalloc_to_pfn(virt);
1417 else
1418 entry->pfn = page_to_pfn(virt_to_page(virt));
1419
c2bbf9d1 1420 add_dma_entry(entry, attrs);
6bfd4498 1421}
6bfd4498
JR
1422
1423void debug_dma_free_coherent(struct device *dev, size_t size,
479623fd 1424 void *virt, dma_addr_t dma_addr)
6bfd4498
JR
1425{
1426 struct dma_debug_entry ref = {
1427 .type = dma_debug_coherent,
1428 .dev = dev,
e57d0552 1429 .offset = offset_in_page(virt),
479623fd 1430 .dev_addr = dma_addr,
6bfd4498
JR
1431 .size = size,
1432 .direction = DMA_BIDIRECTIONAL,
1433 };
1434
3aaabbf1 1435 /* handle vmalloc and linear addresses */
af1da686 1436 if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
3aaabbf1
MC
1437 return;
1438
1439 if (is_vmalloc_addr(virt))
1440 ref.pfn = vmalloc_to_pfn(virt);
1441 else
1442 ref.pfn = page_to_pfn(virt_to_page(virt));
1443
01ce18b3 1444 if (unlikely(dma_debug_disabled()))
6bfd4498
JR
1445 return;
1446
1447 check_unmap(&ref);
1448}
6bfd4498 1449
0e74b34d 1450void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
c2bbf9d1
HM
1451 int direction, dma_addr_t dma_addr,
1452 unsigned long attrs)
0e74b34d
NS
1453{
1454 struct dma_debug_entry *entry;
1455
1456 if (unlikely(dma_debug_disabled()))
1457 return;
1458
1459 entry = dma_entry_alloc();
1460 if (!entry)
1461 return;
1462
1463 entry->type = dma_debug_resource;
1464 entry->dev = dev;
2e0cc304 1465 entry->pfn = PHYS_PFN(addr);
0e74b34d
NS
1466 entry->offset = offset_in_page(addr);
1467 entry->size = size;
1468 entry->dev_addr = dma_addr;
1469 entry->direction = direction;
1470 entry->map_err_type = MAP_ERR_NOT_CHECKED;
1471
c2bbf9d1 1472 add_dma_entry(entry, attrs);
0e74b34d 1473}
0e74b34d
NS
1474
1475void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1476 size_t size, int direction)
1477{
1478 struct dma_debug_entry ref = {
1479 .type = dma_debug_resource,
1480 .dev = dev,
1481 .dev_addr = dma_addr,
1482 .size = size,
1483 .direction = direction,
1484 };
1485
1486 if (unlikely(dma_debug_disabled()))
1487 return;
1488
1489 check_unmap(&ref);
1490}
0e74b34d 1491
b9d2317e
JR
1492void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1493 size_t size, int direction)
1494{
aa010efb
JR
1495 struct dma_debug_entry ref;
1496
01ce18b3 1497 if (unlikely(dma_debug_disabled()))
b9d2317e
JR
1498 return;
1499
aa010efb
JR
1500 ref.type = dma_debug_single;
1501 ref.dev = dev;
1502 ref.dev_addr = dma_handle;
1503 ref.size = size;
1504 ref.direction = direction;
1505 ref.sg_call_ents = 0;
1506
1507 check_sync(dev, &ref, true);
b9d2317e 1508}
b9d2317e
JR
1509
1510void debug_dma_sync_single_for_device(struct device *dev,
1511 dma_addr_t dma_handle, size_t size,
1512 int direction)
1513{
aa010efb
JR
1514 struct dma_debug_entry ref;
1515
01ce18b3 1516 if (unlikely(dma_debug_disabled()))
b9d2317e
JR
1517 return;
1518
aa010efb
JR
1519 ref.type = dma_debug_single;
1520 ref.dev = dev;
1521 ref.dev_addr = dma_handle;
1522 ref.size = size;
1523 ref.direction = direction;
1524 ref.sg_call_ents = 0;
1525
1526 check_sync(dev, &ref, false);
b9d2317e 1527}
b9d2317e 1528
a31fba5d
JR
1529void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1530 int nelems, int direction)
1531{
1532 struct scatterlist *s;
88f3907f 1533 int mapped_ents = 0, i;
a31fba5d 1534
01ce18b3 1535 if (unlikely(dma_debug_disabled()))
a31fba5d
JR
1536 return;
1537
1538 for_each_sg(sg, s, nelems, i) {
aa010efb
JR
1539
1540 struct dma_debug_entry ref = {
1541 .type = dma_debug_sg,
1542 .dev = dev,
0abdd7a8
DW
1543 .pfn = page_to_pfn(sg_page(s)),
1544 .offset = s->offset,
aa010efb
JR
1545 .dev_addr = sg_dma_address(s),
1546 .size = sg_dma_len(s),
1547 .direction = direction,
1548 .sg_call_ents = nelems,
1549 };
1550
88f3907f 1551 if (!i)
aa010efb 1552 mapped_ents = get_nr_mapped_entries(dev, &ref);
88f3907f
FT
1553
1554 if (i >= mapped_ents)
1555 break;
1556
aa010efb 1557 check_sync(dev, &ref, true);
a31fba5d
JR
1558 }
1559}
a31fba5d
JR
1560
1561void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1562 int nelems, int direction)
1563{
1564 struct scatterlist *s;
88f3907f 1565 int mapped_ents = 0, i;
a31fba5d 1566
01ce18b3 1567 if (unlikely(dma_debug_disabled()))
a31fba5d
JR
1568 return;
1569
1570 for_each_sg(sg, s, nelems, i) {
aa010efb
JR
1571
1572 struct dma_debug_entry ref = {
1573 .type = dma_debug_sg,
1574 .dev = dev,
0abdd7a8
DW
1575 .pfn = page_to_pfn(sg_page(s)),
1576 .offset = s->offset,
aa010efb
JR
1577 .dev_addr = sg_dma_address(s),
1578 .size = sg_dma_len(s),
1579 .direction = direction,
1580 .sg_call_ents = nelems,
1581 };
88f3907f 1582 if (!i)
aa010efb 1583 mapped_ents = get_nr_mapped_entries(dev, &ref);
88f3907f
FT
1584
1585 if (i >= mapped_ents)
1586 break;
1587
aa010efb 1588 check_sync(dev, &ref, false);
a31fba5d
JR
1589 }
1590}
a31fba5d 1591
1745de5e
JR
1592static int __init dma_debug_driver_setup(char *str)
1593{
1594 int i;
1595
1596 for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1597 current_driver_name[i] = *str;
1598 if (*str == 0)
1599 break;
1600 }
1601
1602 if (current_driver_name[0])
f737b095 1603 pr_info("enable driver filter for driver [%s]\n",
e7ed70ee 1604 current_driver_name);
1745de5e
JR
1605
1606
1607 return 1;
1608}
1609__setup("dma_debug_driver=", dma_debug_driver_setup);